[CPU] [FP8 SDPA] Enable FP8 SDPA pattern match (#3076)

* [FP8 SDPA] Enable fp8 sdpa pattern match

Author: Valentine233

test/prototype/inductor/test_qsdpa_fusion.py

@@ -18,26 +18,184 @@
 from torchao.utils import torch_version_at_least
 
 
-class SelfAttnLikeModule(torch.nn.Module):
+def qdq(input, scale):
+    dtype = input.dtype
+    q_input = torch.ops.torchao.quantize_affine_float8_non_decomposed.default(
+        input,
+        torch.tensor([scale]),
+        torch.float8_e4m3fn,
+    )
+    dq_input = torch.ops.torchao.dequantize_affine_float8_non_decomposed.default(
+        q_input,
+        torch.tensor([scale]),
+        dtype,
+    )
+    return dq_input
+
+
+def fp8_convert_(model):
+    def generate_model_info(model):
+        from collections import namedtuple
+
+        mod_inst_info = namedtuple("ModInstInfo", ["name", "parent"])
+        parent_child_mod_dict = {}
+
+        def create_mod_info_recursion(parent):
+            for name, mod in parent.named_children():
+                parent_child_mod_dict[mod] = mod_inst_info(name=name, parent=parent)
+                create_mod_info_recursion(mod)
+
+        create_mod_info_recursion(model)
+        return parent_child_mod_dict
+
+    parent_child_mod_dict = generate_model_info(model)
+    for name, mod in model.named_modules():
+        mod_type_str = mod.__class__.__name__
+        if mod_type_str not in [
+            "Linear",
+            "SDPA",
+        ]:
+            continue
+        if mod_type_str == "Linear":
+            param = mod.weight
+            xmax = torch.max(param)
+            weight_scale = xmax / torch.finfo(torch.float8_e4m3fn).max
+            mod.weight_scale = weight_scale
+            q_param = torch.clamp(
+                (param / weight_scale),
+                torch.finfo(torch.float8_e4m3fn).min,
+                torch.finfo(torch.float8_e4m3fn).max,
+            ).to(torch.float8_e4m3fn)
+            mod.weight.data = q_param
+            patched_mod = FP8QDQLinear(mod.in_features, mod.out_features, False)
+            patched_mod.bias = mod.bias
+            patched_mod.weight_scale = weight_scale.item()
+            patched_mod.weight.data = q_param
+        else:
+            patched_mod = FP8QDQSDPA()
+            patched_mod.__dict__.update(mod.__dict__)
+            patched_mod.transpose_for_scores = mod.transpose_for_scores
+
+            patched_mod.q_out_scale = (
+                patched_mod.q_out_scale / torch.finfo(torch.float8_e4m3fn).max
+            )
+            patched_mod.k_out_scale = (
+                patched_mod.k_out_scale / torch.finfo(torch.float8_e4m3fn).max
+            )
+            patched_mod.attn_weights_scale = (
+                patched_mod.attn_weights_scale / torch.finfo(torch.float8_e4m3fn).max
+            )
+            patched_mod.v_out_scale = (
+                patched_mod.v_out_scale / torch.finfo(torch.float8_e4m3fn).max
+            )
+            patched_mod.qk_out_scale = (
+                patched_mod.qk_out_scale / torch.finfo(torch.float8_e4m3fn).max
+            )
+            patched_mod.attn_out_scale = (
+                patched_mod.attn_out_scale / torch.finfo(torch.float8_e4m3fn).max
+            )
+
+        parent = parent_child_mod_dict[mod].parent
+        name = parent_child_mod_dict[mod].name
+        setattr(parent, name, patched_mod)
+    model.eval()
+    return model
+
+
+class FP8QDQLinear(torch.nn.Module):
+    def __init__(self, in_features, out_features, has_bias):
+        super().__init__()
+        self.qtype = torch.float8_e4m3fn
+        self.weight = torch.randn((out_features, in_features)).to(self.qtype)
+        self.weight_scale = 2.0
+        self.scale = 2.0
+        self.bias = None
+        if has_bias:
+            self.bias = torch.randn((out_features,))
+
+    def forward(self, input):
+        weight = torch.ops.torchao.dequantize_affine_float8_non_decomposed.default(
+            tensor=self.weight.data,
+            scale=torch.tensor([self.weight_scale]),
+            output_dtype=torch.float,
+        )
+
+        q_input = torch.ops.torchao.quantize_affine_float8_non_decomposed.default(
+            tensor=input,
+            scale=torch.tensor([self.scale]),
+            float8_dtype=self.qtype,
+        )
+        dq_input = torch.ops.torchao.dequantize_affine_float8_non_decomposed.default(
+            tensor=q_input,
+            scale=torch.tensor([self.scale]),
+            output_dtype=torch.float,
+        )
+
+        out = torch.nn.functional.linear(dq_input, weight, self.bias)
+        return out
+
+
+class FP8QDQSDPA(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.q_out_scale = 1.5
+        self.k_out_scale = 1.5
+        self.attn_weights_scale = 1.5
+        self.v_out_scale = 1.5
+        self.attn_out_scale = 1.5
+        self.qk_out_scale = 1.5
+
+    def forward(self, q, k, v, mask):
+        key = self.transpose_for_scores(q)
+        value = self.transpose_for_scores(k)
+        query = self.transpose_for_scores(v)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        query_qdq = qdq(query, self.q_out_scale)
+        key_qdq = qdq(key.transpose(-1, -2), self.k_out_scale)
+        attn_weights = torch.matmul(query_qdq, key_qdq) / (self.input_dim**0.5)
+
+        # Normalize the attention scores to probabilities.
+        attn_weights = torch.nn.functional.softmax(
+            attn_weights, dim=-1, dtype=torch.float32
+        ).to(query.dtype)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        dropout = 0.0 if not self.training else self.dropout_prob
+        attn_weights = torch.nn.functional.dropout(
+            attn_weights, p=dropout, training=self.training
+        )
+
+        # Mask heads if we want to
+        if mask is not None:
+            attn_weights = attn_weights + mask
+
+        value_qdq = qdq(value, self.v_out_scale)
+        attn_weights_qdq = qdq(attn_weights, self.attn_weights_scale)
+        attn_output = torch.matmul(attn_weights_qdq, value_qdq)
+        attn_output = attn_output.transpose(1, 2).contiguous()
+
+        new_context_layer_shape = attn_output.size()[:-2] + (self.all_head_size,)
+        attn_output = attn_output.reshape(new_context_layer_shape)
+
+        return attn_output
+
+
+class SDPA(torch.nn.Module):
     def __init__(
         self,
         input_dim,
         has_mask,
-        num_attention_heads=None,
-        attention_head_size=None,
+        num_attention_heads,
+        attention_head_size,
     ) -> None:
         super().__init__()
         self.input_dim = input_dim
-        self.q_proj = torch.nn.Linear(input_dim, input_dim, bias=False)
-        self.k_proj = torch.nn.Linear(input_dim, input_dim, bias=False)
-        self.v_proj = torch.nn.Linear(input_dim, input_dim, bias=False)
         self.softmax = torch.nn.Softmax(dim=-1)
-        assert num_attention_heads is not None
-        assert attention_head_size is not None
         self.num_attention_heads = num_attention_heads
         self.attention_head_size = attention_head_size
         self.all_head_size = self.num_attention_heads * self.attention_head_size
-        self.dense = torch.nn.Linear(self.all_head_size, self.all_head_size)
         self.dropout = torch.nn.Dropout(0)
         self.has_mask = has_mask
 
@@ -49,10 +207,7 @@ def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
         x = x.view(new_x_shape)
         return x.permute([0, 2, 1, 3])
 
-    def forward(self, x, mask):
-        q = self.q_proj(x)
-        k = self.k_proj(x)
-        v = self.v_proj(x)
+    def forward(self, q, k, v, mask):
         q = self.transpose_for_scores(q)
         k = self.transpose_for_scores(k)
         v = self.transpose_for_scores(v)
@@ -63,9 +218,38 @@ def forward(self, x, mask):
         attention = self.dropout(attention)
         context_layer = torch.matmul(attention, v)
         context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
-        context_layer = context_layer.view(
-            context_layer.size()[:-2] + (self.all_head_size,)
+        return context_layer.reshape(context_layer.size()[:-2] + (self.all_head_size,))
+
+
+class MHAModule(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim,
+        has_mask,
+        num_attention_heads,
+        attention_head_size,
+    ) -> None:
+        super().__init__()
+        self.input_dim = input_dim
+        self.q_proj = torch.nn.Linear(input_dim, input_dim, bias=False)
+        self.k_proj = torch.nn.Linear(input_dim, input_dim, bias=False)
+        self.v_proj = torch.nn.Linear(input_dim, input_dim, bias=False)
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_size = attention_head_size
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.dense = torch.nn.Linear(self.all_head_size, self.all_head_size)
+        self.attn_mod = SDPA(
+            input_dim,
+            has_mask,
+            num_attention_heads,
+            attention_head_size,
         )
+
+    def forward(self, x, mask):
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+        context_layer = self.attn_mod(q, k, v, mask)
         return self.dense(context_layer)
 
 
@@ -158,7 +342,7 @@ def _check_common(
         reason="cpp kernels not built",
     )
     @config.patch({"freezing": True})
-    def _test_qsdpa_rewriter(self):
+    def _test_int8_sdpa_rewriter(self):
         import torchao.quantization.pt2e.quantizer.x86_inductor_quantizer as xiq
         from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e
         from torchao.quantization.pt2e.quantizer.x86_inductor_quantizer import (
@@ -171,7 +355,7 @@ def _test_qsdpa_rewriter(self):
             [torch.float32, torch.bfloat16], [True, False], [56, 1]
         ):
             seqlen, numhead, headsize = 197, 16, 64
-            mod = SelfAttnLikeModule(
+            mod = MHAModule(
                 input_dim=headsize * numhead,
                 has_mask=has_mask,
                 num_attention_heads=numhead,
@@ -204,6 +388,51 @@ def _test_qsdpa_rewriter(self):
                 prepare_model(*inputs)
                 convert_model = convert_pt2e(prepare_model)
                 torchao.quantization.pt2e.move_exported_model_to_eval(convert_model)
+
+                self._check_common(
+                    convert_model, args1=inputs, check_train=False, atol=1.0
+                )
+
+    @skipIfRocm
+    @unittest.skipIf(
+        not torch_version_at_least("2.7.0"),
+        reason="qsdpa requires torch 2.7 or later",
+    )
+    @unittest.skipIf(
+        "CPU" not in torch._C._dispatch_dump("torchao::qscaled_dot_product"),
+        reason="cpp kernels not built",
+    )
+    @config.patch({"freezing": True})
+    def _test_fp8_sdpa_rewriter(self):
+        import torchao.quantization.pt2e.quantizer.x86_inductor_quantizer as xiq  # noqa: F401
+
+        # pattern is different for bs=1
+        torch.manual_seed(1234)
+        for dtype, bs in itertools.product([torch.float32, torch.bfloat16], [56, 1]):
+            seqlen, numhead, headsize = 197, 16, 64
+            mod = MHAModule(
+                input_dim=headsize * numhead,
+                has_mask=False,
+                num_attention_heads=numhead,
+                attention_head_size=headsize,
+            ).eval()
+            inputs = (
+                torch.randn(
+                    (bs, seqlen, headsize * numhead), device=self.device, dtype=dtype
+                ),
+                None,
+            )
+            enable_autocast = dtype == torch.bfloat16
+            with (
+                torch.no_grad(),
+                torch.amp.autocast(
+                    self.device, enabled=enable_autocast, dtype=torch.bfloat16
+                ),
+                config.patch(post_grad_custom_pre_pass=custom_pass),
+            ):
+                _qsdpa_init()
+                convert_model = fp8_convert_(mod)
+
                 self._check_common(
                     convert_model, args1=inputs, check_train=False, atol=1.0
                 )
@@ -213,7 +442,12 @@ def _test_qsdpa_rewriter(self):
 
     class SDPAPatternRewriterCpuTests(TestSDPAPatternRewriterTemplate):
         device = "cpu"
-        test_qsdpa_rewriter_cpu = TestSDPAPatternRewriterTemplate._test_qsdpa_rewriter
+        test_int8_sdpa_rewriter_cpu = (
+            TestSDPAPatternRewriterTemplate._test_int8_sdpa_rewriter
+        )
+        test_fp8_sdpa_rewriter_cpu = (
+            TestSDPAPatternRewriterTemplate._test_fp8_sdpa_rewriter
+        )
 
 
 if __name__ == "__main__":